DBSCAN does not make outliers clear

[lilyminium]

Expected behavior

DBSCAN is a clustering method that can identify outliers. I expect these outliers to be clearly indicated in some way. I also expect that outliers are treated properly in similarity measures.

Actual behavior

The method implemented in MDAnalysis makes the outlier group (label==-1) look like an actual cluster (labels start from 0). (although this doesn't ultimately matter, as encode_centroid_info drops label information anyway)

https://github.com/MDAnalysis/mdanalysis/blob/9bcf6f4c118e1ea137e8514bd60cbd1cd1972062/package/MDAnalysis/analysis/encore/clustering/ClusteringMethod.py#L300-L306

Also, calling the first frame in the cluster the centroid, and not mentioning this very clearly in the docs seems like a bad idea. This also gives the outlier group a centroid.

Finally, ClusterCollection does not keep the cluster labels. This makes it hard to look for special (i.e. negative) cluster labels.

Currently version of MDAnalysis

• Which version are you using? (run python -c "import MDAnalysis as mda; print(mda.__version__)") 0.20.2-dev
• Which version of Python (python -V)?
• Which operating system?

Possible fix

Easy option

• Don't alter DBSCAN's output
• Add a warning and note in the docs that the "centroid" is the first frame of that cluster
• Figure out first frame in the outlier group (which becomes the "centroid" in ClusterCollection) and add a warning that it's not a real cluster

More work option

• Reconstruct the ClusterCollection class with a more intuitive interface
  • each cluster should not require a centroid
  • each cluster should retain its label from scipy
  • should be able to label a "cluster" as outliers
  • it would be nice to link the cluster members to frames of the universes in the ensemble

[lilyminium]

This also results in issues for ensemble similarity analysis. The outlier "cluster" is treated like a real cluster. Therefore, if a conformation in trajectory A is in the outlier cluster and a conformation in trajectory B is in the outlier cluster, it is treated as a point of similarity -- in reality these conformations should be unrelated.

[mtiberti]

We will update the documentation and code to add a warning when DBScan is being used - so that users are aware of this issue